It appears that Nvidia is ready to talk a bit more about its upcoming next-gen Pascal GPU As the company unveiled a bit more information during Graphics Technology Conference held in Japan.

According to details published by VR World, the Pascal GPU is designed to offer double the performance per Watt compared to the current Maxwell architecture. Most of the performance gain will be coming from stacked HBM2 (High Bandwidth Memory gen. 2). According to details, the high-end Pascal GPU part will feature four 4-gigabyte HBM2 stacks, adding up to 16GB of memory with a combined memory bandwidth of 1TB/s. Internally, bandwidth can reach as high as 2TB/s.

The architecture and the GPU itself supports up to 32GB of memory so it is possible that Nvidia's Quadro and Tesla graphics card could peak at 32GB.

The rest of the revealed details suggest that Nvidia's Pascal GPU architecture will be based on the 16nm FinFET manufacturing process. It will feature a new interconnect called the NVLink, which will change the way that dual-GPU graphics cards are made.

While current dual-GPU graphics cards are pretty much two GPUs on the same PCB, where they share PCI-Express bandwidth via bridge-chip and internal SLI bridge, the NVLink will feature an 80GB/s bi-directional data path, letting each GPU directly address memory controlled by the other. This will significantly improve the performance and memory management in games developer on next-gen APIs, like DirectX 12 and Vulkan.

According to current and earlier reports, first graphics cards based on Pascal GPU architecture should be ready in the first half of 2016.



Source: via Techpowerup.com.